AI 'solved' famous maths problems. Read the asterisks.

Two claims, one week. OpenAI: an internal reasoning model disproved Erdős's 1946 unit-distance conjecture. DeepMind: AlphaProof Nexus solved nine open Erdős problems plus 44 conjectures. The dunk-tweet version — 'Google beat OpenAI nine to one' — is fun and almost entirely beside the point, because the two results aren't even the same kind of thing. One is a single flashy disproof awaiting review; the other is a stack of formally certified proofs. Counting them against each other is a category error dressed up as a leaderboard.

OpenAI's proof is human-checked and not yet through formal peer review. Real mathematicians — Timothy Gowers among them — read it, which is genuine evidence, but expert eyes have missed subtle gaps before; that's the whole reason peer review exists in the first place. DeepMind's nine, by contrast, are machine-verified in Lean. That is the genuinely impressive part — and it also means the win comes from a tight generate-then-verify loop, not a model soloing genius in one heroic leap. The model writes proof steps in Lean's formal language; the compiler throws out anything that doesn't hold and feeds the error back for another try. The cleverness is real; so is the scaffolding holding it up, and honest coverage names both.

Put the two side by side and the 'who won' question dissolves:

The scoreboard crowd is comparing a flashy single disproof against a pile of certified ones. Different sports.

Here's the part the cynics get wrong, though. The fact that DeepMind leaned on Lean isn't a gotcha — it's the innovation. A loop where 'AI proposes, formal system filters' works only because, in maths, 'correct' is mechanically checkable. That's rare. Most of the world — strategy, taste, law, the messy middle of science — has no Lean to keep the model honest. So the achievement is sharply bounded: brilliant inside a checkable box, untested outside it. AlphaProof Nexus also builds on DeepMind's earlier olympiad prover, which means this isn't a system that woke up able to do research maths; it's years of narrow, deliberate engineering pointed at exactly the kind of problem it can verify. Genuinely hard, genuinely narrow — both at once.

The tell is that the people closest to it are the most careful. Hassabis said straight out the system is 'still not AGI'. When the vendor undersells and the timeline oversells, trust the vendor. AI is now a real instrument for parts of maths — narrow parts, with a human or a checker holding the pen — and pretending otherwise just sets up the next disappointment cycle. The hype merchants and the dismissers are both wrong, in mirror-image ways; the accurate read sits unglamorously in the middle.

So file it correctly. A real milestone, a narrow one, and a preview of how AI actually gets useful in serious domains: not by being trusted, but by being checked. The headline 'AI solves famous maths' is true. The implied 'AI is now a mathematician' is not. The machinery in between is the only part worth screenshotting.